1/x

Who's ready for an election thread and a (statistically literate) poll analysis & update?

There's an OBSESSION with "what the polls missed" in 2016. If you follow me, you'd know: the polls weren't wrong - people just read them wrong.

2/x

That's part of why @FiveThirtyEight's statistically invalid analysis of

"poll margin - election margin = poll error"

Is so damaging. Not only is it logically and statistically invalid, it leads them/the public to believe the POLLS were wrong when that's likely not true.

3/x

In 2016, Hillary had a decent lead, IF YOU ONLY LOOKED AT MARGIN

But you'll note - this is important -IN NO SWING STATE DID SHE POLL ABOVE 47%

Compare that to 2020

Biden is polling at ABOVE 49% in MI, PA, WI, NH, MN, ME, FL, NC, NV (and above 50% in the first 6 of those)

4/x

Many of you already know this. But I'm going to say it for the 6,969th time:

Look at those yellow "undecided" bars.

Polls tell us NOTHING about how they'll eventually vote, nor do they try to.

So when we talk about what polls *do* tell us, look at the actual numbers.

5/x Here are the 2016 pre-election polls compared to the results.

You'll note the large yellow bars - undecideds - largely stacked onto Trump's side.

Is that a polling error? Nope. To call it a polling error is, plainly, wrong.

So, what does this mean for 2020?

6/x

Here's the updated Trump-leaning forecast

Biden polling at or near 50% is flat out insurmountable, no matter how you put it

The "polling error" (as it's wrongly called) came from undecideds, NOT from overestimating a candidate's support

So where does Trump make up ground?

7/x

In short, he doesn't. Even in this undecideds-generous forecast favoring Trump, he's too close to call IN GEORGIA.

He holds onto Texas, Iowa, and Ohio, but loses by an easily called margin in MI, PA, WI, NC, MN (and probably FL, AZ, and NC too)

8/x

Now, I'm not saying polls CAN'T be wrong.

If Biden comes in 2 or 3 points lower than his poll avg, that is almost certainly a polling error. However, largest error of this type (election number below poll number) I can find since 2012 is 1.7%. And it wasn't across the board

9/x

But in order for Trump to HAVE A CHANCE (not win, but simply have a chance) Biden would need to lose, or for there to be a systemic error of 2-3 points in FL, NC, WI, MI, PA..

Again, not saying this can't happen, but this type of polling error is NOT what happened in 2016.

10/x

Polls ONLY tell us:

1) Preferences of decided voters
2) HOW MANY undecided voters

That's it.

A polling error IS NOT undecideds breaking 50/50 vs 80/20. Polls don't try to predict that

Wrongly projecting where undecideds will go is a FORECASTING ERROR ( @FiveThirtyEight)

11/x

Every good forecast starts with polling. Polls are the raw data that we forecast from. But it's VERY IMPORTANT we don't assign meaning to them that we shouldn't: namely, using poll margin as a guide to election margin

Doing so literally assumes undecideds will break 50/50

12/x

If I said "we will assume that undecideds break 50/50" as an axiom to evaluating poll validity, you'd likely be very quick to say "um, that's not right"

Yet, that's exactly what 538 does with their "poll accuracy" formula.

46-40 doesn't imply 53-47.

13/x

Now, I'm hard on @FiveThirtyEight because, frankly, they're wrong and deserve to be called on it. But it's important to mention:

Their MODELS are rock solid and they're the best poll aggregators I've seen. ( @jhkersting and @gelliottmorris are also good, fwiw)

14/x

The problem is, being the most respected name in poll data, they've done MAJOR disservice to #electiontwitter (and the general public) in terms of explaining, simply, what polls do

Poll margin vs Election margin isn't a "polling error" but a lot of people assume that now

15/x

Part 2 of my forecast, if undecideds break in a way that favors Biden. I think there's more evidence for this than the opposite (demographics of undecideds).